Manual CCS site characterisation runs four to seven years from basin study to first injection permit — too slow to close the IEA's 1.7 Gt/year storage gap by 2030. A five-agent orchestrated loop compresses that screening step from months to minutes, with no loss of auditability.
Abstract
Carbon capture and storage at climate-relevant scale requires a 30-fold expansion of annual injected volumes by 2030 [1]. The geological capacity exists — IPCC saline aquifer estimates run to 1,000–10,000 Gt [2] — but the conventional site-characterisation workflow does not. A basin study, site-specific characterisation, and permit application typically consume four to seven years before a single tonne of CO₂ is injected.
We implement a five-agent orchestrated loop (DataIngestor → ReservoirScreener → CaprockAnalyst → InjectivityEstimator → ReportWriter) that screens well candidates against dual reservoir-caprock criteria with automatic fallback re-routing when a primary interval fails caprock integrity. Applied to a 10-well synthetic North Sea / Gulf Coast dataset, the loop ranks six candidates by a Darcy-simplified injectivity index, eliminates four on screening criteria, and triggers one orchestrator fallback — all in under two seconds on commodity hardware.
The headline result is not the runtime. It is that basin-wide legacy-well screening — typically months of manual cross-disciplinary work — is compressible to minutes with full per-criterion auditability, provided the data-normalisation layer is solved first. This note documents the architecture, the screening thresholds, the defining equation, and the failure modes that block production deployment.
What this note shows
- Five-agent loop replaces a months-long manual workflow with a sub-2-second screening pass on 10 wells.
- Orchestrator handles caprock failures by re-routing to deeper intervals — no human handoff required.
- Top candidate (DEMO-04, Bunter Sandstone analogue) ranks at 187 m³/d/bar injectivity index.
- Synthetic data establishes architectural viability; production deployment is gated by depth-datum, unit, and lithology-flag normalisation.
Background
The IEA's net-zero pathway requires roughly 1.7 Gt of CO₂ captured and stored annually by 2030, up from approximately 50 Mt today [1]. The Global CCS Institute's 2025 status report counts on the order of 40 large-scale operational projects globally — an order of magnitude below the deployment trajectory the IEA target implies [3]. The constraint is not capture technology, and it is not geology. It is screening velocity.
screening velocityThe rate at which candidate sites can be evaluated against reservoir + caprock + injectivity criteria. The binding constraint on CCS deployment at climate scale — not subsurface capacity.A conventional site assessment runs in four sequential phases: a basin study (6–12 months), a site-specific characterisation (12–24 months), a permit application (12–36 months), and finally first injection. The same disciplines — petrophysics, structural geology, reservoir engineering, geomechanics — that an upstream asset team mobilises for an exploration well are mobilised again for every candidate storage site, with the same manual handoffs between teams.
| Metric | Current state | Net-zero target (2030) |
|---|---|---|
| Annual CO₂ stored (Gt) | ~0.05 | ~1.7 |
| Global CCS projects (large-scale) | ~40 | ~200+ |
| Average screening-to-permit (years) | 4–7 | < 2 |
| Saline aquifer capacity (Gt, IPCC) | 1,000–10,000 | — |
Recent industrial deployments suggest the workflow is compressible. ADNOC's ENERGYai sovereign LLM, trained on decades of subsurface data and deployed across 28 producing fields, reports a 75% acceleration in geological model build under a $340M three-year contract [4]. The data substrate is the same as CCS — well logs, seismic, pressure records, core — and so is the bottleneck: manual orchestration across siloed teams. The 2024 frame for AI in CCS was a chatbot answering questions about porosity thresholds. The 2026 frame is an autonomous agent loop that perceives the basin data, reasons over screening criteria, acts against a well database, and hands a ranked candidate list to the geoscientist for adjudication.
Public basin datasets are converging on a form the loop can consume. The USGS CarbonSAFE programme is publishing characterisation data for U.S. basins [5], and the UK North Sea Transition Authority and Norway's NPD have moved furthest on digitising storage-permit submissions. The architectural question is whether the screening step itself can be expressed as a composable agent graph rather than a linear pipeline.
Method
The architecture is a five-agent orchestrated loop. Each agent exposes explicit reason() and act() methods and operates on a shared stateful WellRecord object that carries depth, porosity, permeability, net pay, temperature, caprock thickness, continuity flag, and fault proximity. The orchestrator routes state between agents and handles fallback re-routing when an early-stage agent fails on a recoverable criterion.
Five-agent orchestrated loop
Each agent has explicit reason() / act(); orchestrator handles fallback re-routing
Agent 1 — DataIngestor
Reads well headers, normalises datum, flags missing fields.
Agent 2 — ReservoirScreener
Depth 800–3,000 m, φ > 10%, k > 10 mD, T < 120°C.
Agent 3 — CaprockAnalyst
Seal ≥ 50 m, continuous, no fault within 2 km.
Agent 4 — InjectivityEstimator
Darcy-simplified k·h / μ_CO₂ index, m³/d/bar.
Agent 5 — ReportWriter
Ranked JSON + human-readable summary.
Agent 1 — DataIngestor — reads well header data, normalises depth datum, and flags missing mandatory fields. Wells failing the required-field check are skipped, not silently defaulted. Agent 2 — ReservoirScreener — applies CCS suitability thresholds: depth 800–3,000 m (the supercritical CO₂ window), porosity > 10%, permeability > 10 mD, temperature < 120°C. It fails fast on show-stoppers and emits a structured reason string. Agent 3 — CaprockAnalyst — checks shale or evaporite seal integrity above the candidate reservoir: minimum thickness 50 m, continuity flag, and no mapped fault intersections within 2 km. Agent 4 — InjectivityEstimator — computes a Darcy-simplified CO₂ injectivity index. Agent 5 — ReportWriter — assembles the ranked candidate list with pass/fail flags per criterion and writes structured JSON alongside a human-readable summary.
The injectivity index is the simplified Darcy form, CO₂-corrected for reservoir-condition viscosity:
I = \frac{k \cdot h}{\mu_{\mathrm{CO_2}} \cdot S}where k is permeability in mD, h is net pay thickness in m, μ_CO₂ ≈ 0.05 cP is CO₂ viscosity at reservoir conditions, and S = 481 mD·m·cP / (m³/d/bar) is a Darcy-simplified unit-normalisation constant. Output units are m³/d/bar. The formula deliberately omits skin factor, drainage radius, relative permeability, and PVT variation with depth — a production screener at Agent 4 would substitute a full numerical reservoir model.
The agentic distinction versus a linear pipeline lies in the orchestrator's fallback behaviour. When Agent 3 reports caprock failure on the primary interval, the orchestrator does not discard the well. Instead, it consults a deeper-intervals table, re-instantiates the WellRecord with the alternative interval's parameters, and re-routes through Agents 2 and 3 again. The geoscientist adjudicates outputs, not steps.
Why agentic, not pipelined
The synthetic dataset comprises 10 wells with North Sea / Gulf Coast analogue parameters spanning the design envelope: depth 620–3,200 m, porosity 6–22%, permeability 8–150 mD, caprock thickness 35–90 m. The distribution is deliberately mixed to exercise every branch of the agent graph — clear passes, single-criterion fails, the supercritical-window boundary on both sides, and one well configured to fail caprock on interval 1 and reservoir on interval 2 so the orchestrator's fallback path is invoked exactly once.
Results
Of the 10 candidate wells, six passed both reservoir and caprock screening and were ranked by injectivity index. Four were eliminated on hard thresholds, and one (DEMO-09) triggered the orchestrator fallback: its primary interval failed caprock on seal thickness (35 m, below the 50 m minimum), the orchestrator re-routed to a deeper interval at 2,050 m, and that interval failed reservoir screening on porosity (9%, below the 10% threshold). The well exits with a PASS / FAIL pair and a fallback note in the audit trail.
10-well synthetic screening pass
Candidates ranked
Eliminated on thresholds
Orchestrator fallback triggered
Total pipeline runtime
The top-ranked candidate, DEMO-04 (Bunter Sandstone analogue, 1,450 m depth, 22% porosity, 150 mD permeability, 30 m net pay), returned an injectivity index of 187 m³/d/bar. The lowest-ranked passing candidate, DEMO-08 (near the temperature limit at 118°C with 28 mD permeability and 16 m net pay), returned 19 m³/d/bar — a tenfold spread across passing candidates, which is the point of ranking rather than binary screening.
Total runtime for the 10-well pass — including data ingestion, screening, caprock analysis, the orchestrator fallback, injectivity estimation, ranked-report assembly, and JSON serialisation — was under two seconds on commodity hardware with the Python standard library and no external dependencies.
Failure modes captured by the loop
Porosity 6% < 10% threshold
Depth 620 m below supercritical window
Caprock 35 m → fallback → φ 9% fail
Depth 3,200 m above 3,000 m limit
Extrapolating to basin scale: a national geological survey's legacy database of 2,000 exploration wells would screen in minutes on the same hardware. The geoscientist's adjudication time, not the compute, becomes the bottleneck.
Discussion
The result demonstrates that the screening step of CCS site characterisation is compressible by roughly four orders of magnitude on synthetic data. That compression is conditional. Production deployment is gated not by the agent architecture but by the data-engineering layer beneath it — and the CCS data problem is materially worse than the upstream oil-and-gas baseline because CCS datasets span operator, era, and country.
Three failure modes recur. First, depth-datum chaos: KB, mean sea level, and seabed datums coexist in any multi-well dataset, and Agent 2's 800 m lower bound is meaningless until every well is on a common reference. This is an Agent 0 pre-processor if done systematically, not a two-hour manual fix per well. Second, missing caprock flags: standard LAS files do not carry a shale-flag channel, so Agent 3 needs either a formation-tops table (often absent in early-stage screening) or a real-time lithology classifier on the gamma-ray curve. Third, unit inconsistency: mD versus μm², ft versus m, psi versus bar — each ambiguity is a silent factor-of-10 error in the injectivity estimate. The loop should reject, not silently convert.
The boring layer is the binding layer
The architecture has explicit limitations worth naming. The injectivity formula omits skin factor, drainage radius, relative permeability, and CO₂ PVT variation with depth — a production screener requires a full Darcy or numerical reservoir model at Agent 4. The thresholds are binary, with no uncertainty propagation from log-derived porosity or permeability. The wells are evaluated independently, with no basin map, fault network, or stratigraphic framework — site interference, plume migration, and pressure build-up between adjacent injection sites are not modelled. Pipeline proximity is absent from the ranking, although transport-network constraints (Porthos in Rotterdam, Northern Lights in Norway) materially change site economics. The orchestrator's fallback re-routes to one alternative interval per well, where stacked Rotliegend / Bunter basins would benefit from a depth-range iterator.
Two of the agent objects generalise across the CCS / CCU boundary. The InjectivityEstimator is directly reusable for CO₂-EOR candidate screening — same k·h logic, different target thresholds. The CaprockAnalyst, post-injection, becomes a containment monitor with the same data structures and a shifted decision boundary. The agent objects were designed for that reuse.
The geoscientist's role shifts. With basin-scale screening compressed to minutes, the value-add moves from running the screening to adjudicating ranked outputs, owning the uncertainty quantification on porosity and permeability, and instrumenting the fallback frequency as a data-quality signal — a high re-route rate is the loop telling the asset team that the data layer needs work before the screening output is trustworthy.
The geoscientist adjudicates outputs, not steps. Fallback frequency becomes the data-quality signal.
From manual workflow to orchestrated loop
The compression is most legible as a before-and-after on the same screening task. A conventional basin-wide pass through 10 candidate wells — datum reconciliation, threshold checks, caprock review, injectivity estimation, ranked write-up — runs as a cross-disciplinary exercise measured in weeks. The orchestrated loop runs it as a single forward pass.
Before
weeks
After
< 2 s
References
[1] IEA (2023). Net Zero by 2050 — A Roadmap for the Global Energy Sector (updated). International Energy Agency. Storage capacity targets, §4. https://www.iea.org/reports/net-zero-by-2050
[2] IPCC (2005). Special Report on Carbon Dioxide Capture and Storage. Cambridge University Press. Saline aquifer capacity estimates, Ch. 5. https://www.ipcc.ch/report/carbon-dioxide-capture-and-storage/
[3] Global CCS Institute (2025). Global Status of CCS Report 2025. GCCSI. https://www.globalccsinstitute.com/resources/global-status-report/
[4] ADNOC / AIQ (2025). ENERGYai sovereign LLM deployment — 75% geological model acceleration, 70B parameter model, 28 fields. AIQ communications, March 2025.
[5] USGS CarbonSAFE Programme. Basin characterisation datasets. https://www.usgs.gov/programs/carbonsafe
[6] Northern Lights Project. CO₂ transport and storage infrastructure, Norway. https://northernlightsccs.com
